What is the OpenAI Jalapeno chip? The OpenAI Jalapeno chip is OpenAI's first custom artificial intelligence accelerator, co-developed with Broadcom and unveiled on June 24, 2026, built specifically for large language model inference rather than the general-purpose work that Nvidia GPUs handle. For commercial real estate investors, the news is not really about silicon, it is about data center demand: a chip designed to run AI models faster and at substantially better performance per watt reshapes how much power, space, and capital the next wave of AI buildout requires. That makes Jalapeno directly relevant to anyone underwriting data centers, which have become one of the most consequential property types covered in our best AI tools for commercial real estate investors work.
Key Takeaways
- OpenAI and Broadcom unveiled Jalapeno on June 24, 2026, OpenAI's first custom chip, designed from the ground up for LLM inference rather than training.
- The chip went from design to tape-out in roughly 9 months and is reported to deliver performance per watt substantially better than current state-of-the-art accelerators.
- The partners are targeting gigawatt-scale data centers with Microsoft and other partners, with meaningful ramp expected in 2027 and full scale by 2028.
- For CRE, more efficient inference silicon shifts the question from raw GPU count to power efficiency, location, and the durability of data center demand investors are underwriting.
- Custom silicon deepens tenant and platform concentration risk, which belongs in any data center underwriting alongside power, cooling, and lease term.
What OpenAI and Broadcom Actually Announced
OpenAI and Broadcom announced Jalapeno, a purpose-built inference accelerator that marks OpenAI's move from a pure software company into custom hardware. The chip is engineered specifically for running live AI models, the inference workload that happens every time a user queries a model, as opposed to the training runs that build the models in the first place. Richard Ho, who leads OpenAI's hardware program, said the architecture was optimized around the memory movement, networking, and serving patterns that matter most for frontier models, and engineering samples are already running workloads including a GPT-5 class coding model in the lab.
The development speed is itself notable: the partners took Jalapeno from initial design to manufacturing tape-out in about 9 months, an unusually fast cycle for an advanced semiconductor, aided by using OpenAI's own models in the design process. The die is reticle-scale, roughly 840 square millimeters, and early testing points to performance per watt meaningfully better than today's leading GPUs. Broadcom CEO Hock Tan told CNBC the work would see small prototype deployment late in 2026, ramp in 2027, and go full tilt by the first half of 2028. This sits alongside the broader merchant-silicon shift we covered in Amazon Trainium AI chips and data center CRE.
Inference Versus Training: Why the Distinction Matters for CRE
The inference-versus-training distinction matters because the two workloads have different real estate footprints. Training is concentrated, power-hungry, and latency-tolerant: it favors massive campuses in remote, cheap-power locations where a single site can pull hundreds of megawatts or more. Inference is the opposite in important ways, it is continuous, latency-sensitive, and tied to where users are, which historically pushes some of it toward distributed, better-connected locations closer to population centers. As AI usage shifts from building models to serving billions of daily queries, the balance of demand tilts toward inference.
A chip purpose-built for efficient inference accelerates that tilt. If inference gets cheaper and more power-efficient per query, two things can happen at once: total inference volume rises because more applications become economical, and the power required per unit of useful work falls. For CRE investors, that means the data center demand story is not a single curve. The training megacampus thesis and the distributed inference thesis call for different sites, different power profiles, and different tenants. Underwriting data centers as one undifferentiated asset class misses this, a theme we explore in our analysis of AI data center oversupply risk.
The Power and Efficiency Question
The central CRE question Jalapeno raises is whether more efficient silicon eases or intensifies the power constraint that now dominates data center site selection. Power availability and speed to energization, not connectivity, have become the binding constraints on data center development, with operators racing to secure large blocks of capacity and gigawatt-scale pre-leasing appearing in places like West Texas, a dynamic CBRE tracks across its data center research. A chip with substantially better performance per watt could, in theory, deliver more compute within a fixed power envelope, stretching constrained grids further.
History suggests the opposite is more likely in the near term. When compute gets cheaper and more efficient, demand for it tends to rise faster than efficiency improves, so total power draw keeps climbing even as each chip gets more efficient. The gigawatt-scale data centers OpenAI and Broadcom are explicitly targeting are evidence of that dynamic. For investors, the takeaway is that efficiency gains are unlikely to relieve the power bottleneck soon, so the value of sites with secured, deliverable power remains high. That underwriting reality is consistent with the revenue surge we documented in our piece on the CBRE AI data center boom. CRE investors weighing data center exposure increasingly model these scenarios with The AI Consulting Network.
What Data Center Investors Should Underwrite Now
Data center investors should underwrite Jalapeno as a signal about demand durability and concentration, not as a near-term change to any specific building. The chip will not be at scale until 2027 or 2028, so it changes nothing about cash flows next year. What it does is sharpen three questions that belong in every data center model. First, demand durability: is the project leased to a tenant whose AI buildout is funded and strategic, or to a speculative user that a more efficient compute stack could leave overbuilt? Second, power: are the megawatts secured, deliverable, and priced, because that is the scarce asset regardless of which chip wins.
Third, and increasingly important, concentration risk. The AI infrastructure market is consolidating around a handful of platforms, OpenAI, Nvidia, Amazon, Microsoft, and Google, and custom silicon deepens the ties between a chip, a model provider, and the data centers built to house them. A site purpose-fit for one platform's hardware and power profile is less fungible than a generic facility, which affects re-leasing risk and exit value. Underwriting a long data center hold means stress-testing what happens if the anchor tenant's technology roadmap shifts. For investors who want a structured framework to weigh these factors, The AI Consulting Network helps CRE teams translate AI infrastructure news into underwriting assumptions rather than headlines.
The Bottom Line for CRE
The bottom line is that Jalapeno reinforces, rather than rewrites, the data center thesis: AI demand is real, power is the binding constraint, and the buildout is accelerating toward gigawatt scale. The chip is a reminder that the AI infrastructure stack is maturing and consolidating, which raises both the opportunity and the concentration risk in data center real estate. Investors who treat data centers as a monolithic bet on AI are exposed; those who underwrite the specific demand, power, and tenant profile of each deal are positioned to benefit from the buildout without owning its tail risk. For hands-on help turning announcements like this into a defensible underwriting model, CRE investors can reach out to Avi Hacker, J.D. at The AI Consulting Network.
Frequently Asked Questions
Q: What is the difference between AI training and AI inference chips?
A: Training chips build AI models by processing massive datasets, a concentrated, power-hungry, latency-tolerant workload that favors remote megacampuses. Inference chips like Jalapeno run finished models to answer user queries, a continuous, latency-sensitive workload that is more tied to where users are. The shift toward inference changes the location and power profile of data center demand.
Q: Does the Jalapeno chip change data center demand in 2026?
A: Not immediately. The chip is not expected to reach scale until 2027 or 2028, so it does not affect near-term cash flows. Its significance for CRE is as a signal: it confirms gigawatt-scale AI buildout is accelerating and that power efficiency and platform concentration are becoming central to how data center demand should be underwritten.
Q: Should CRE investors be worried about data center oversupply?
A: The risk is real but uneven. More efficient inference silicon could leave speculative, poorly located facilities overbuilt while well-located projects with secured power and funded, strategic tenants remain in high demand. The defense is underwriting each deal's specific demand durability, power position, and tenant concentration rather than treating data centers as a single AI bet.